FlexDML: High Utilization Configurable Multimode Arithmetic Units Featuring Dual Mode Logic

Dual Mode Logic (DML) enables flexible energy-delay optimization. By setting the design elements in static or dynamic mode, it is possible to significantly improve performance at the cost of a limited increase in energy consumption, and vice versa. In this work, for the first time, we design, fabric...

Full description

Saved in:
Bibliographic Details
Published inIEEE solid-state circuits letters Vol. 6; p. 1
Main Authors Stanger, Inbal, Shavit, Netanel, Taco, Ramiro, Lanuzza, Marco, Yavits, Leonid, Levi, Itamar, Fish, Alexander
Format Journal Article
LanguageEnglish
Published Piscataway IEEE 01.01.2023
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Subjects
Online AccessGet full text

Cover

Loading…
More Information
Summary:Dual Mode Logic (DML) enables flexible energy-delay optimization. By setting the design elements in static or dynamic mode, it is possible to significantly improve performance at the cost of a limited increase in energy consumption, and vice versa. In this work, for the first time, we design, fabricate in a commercial 16nm FinFET process, and evaluate in silicon FlexDML, an adaptively configurable arithmetic unit. FlexDML comprises several DMLBricks, 2×2 and 4×4 DML-based multiplier units, as well as DML-based shifters and adders. FlexDML is capable of performing several arithmetic operations with flexible wordlength, such as N×N-bit, two (N/2)×N-bit or N×(N/2)-bit, four (N/2)×(N/2)-bit and so on multiplications, as well as shifts and inner product calculations. All DMLBricks can be operational at the same time, increasing the hardware utilization to the level unattainable in typical designs that support variable wordlength. FlexDML is the first to support a pipeline multimode, where the DML mode is set separately for each pipeline stage, thus enabling more flexible energy-delay optimization. FlexDML achieves up to 41% performance improvement in dynamic mode, and power savings of about 23% in static mode, compared to CMOS implementation.
ISSN:2573-9603
2573-9603
DOI:10.1109/LSSC.2023.3259102